How Facebook (Meta) Ad Review Works — And What Andromeda Changed

Meta reviews ads with automated systems plus human review, and it keeps reviewing them after they go live — Meta states an ad may not be checked against all policies before it starts delivering, and that all ads are subject to re-review at any time. In late 2024 and 2025 Meta also rebuilt how ads are selected (Andromeda) and ranked (GEM). The part most explainers skip: neither of those is an enforcement system. Here is how review actually works, what triggers rejection in regulated verticals, and what Andromeda and GEM did — and did not — change.

Key takeaways

Review is layered: automated screening plus human escalation, and live ads are re-reviewed after negative feedback or at random.
The rules that reject ads are the Advertising Standards — unchanged by Andromeda and updated on their own schedule.
Andromeda = retrieval (which ads get considered). GEM = ranking (which considered ad wins). Neither writes or enforces policy.
"Andromeda made rejections worse" is advertiser observation, not a Meta statistic — Meta publishes no rejection rates.
The practical shift: creative volume grew, so per-ad checks became per-variant checks.

How Meta ad review works

VERIFIED FACT. Meta describes an ad as subject to review before it can start delivering and to re-review at any time; an ad may not be checked against every policy before it runs (About Ads In Review). Meta says review relies primarily on automated tools and, in some cases, human reviewers:

Automated screening. Systems scan ad copy, creative, and landing-page content against the Advertising Standards, including patterns associated with prior rejections in the same vertical.
Human escalation. Flagged ads, and ads in restricted categories, can be escalated to human reviewers, who evaluate context and can overturn an automated decision.
Continuous re-review. Live ads get pulled back into review after negative feedback — people hiding, blocking, or reporting — or at random for accuracy checks (Why some ads are approved, then rejected).

What triggers a rejection in regulated categories

The Advertising Standards that reject ads are documented separately from delivery systems, and severe or repeated violations escalate from ad-level to account-level enforcement across the business portfolio, ad account, Page, or user account (Common examples of Meta advertising policy violations). The patterns advertisers hit most in health, supplement, and finance verticals are:

Disease, treatment, or clinical-sounding claim language.
Overpromise framing — "guaranteed", "instantly", "100 percent", and similar absolute outcomes.
Personal-attribute wording that implies the reader's health, financial, or other sensitive status.
Before/after imagery and implied transformation narratives told without claim words.
A mismatch between the ad and what the landing page actually says.

What Andromeda actually is

VERIFIED FACT. Meta's engineering post (December 2, 2024) describes Andromeda as a personalized ads retrieval engine: the first stage of the recommendation pipeline, selecting from tens of millions of ad candidates down to a few thousand per person. Documented points (Meta Engineering: Meta Andromeda):

A custom deep neural network co-designed with NVIDIA Grace Hopper Superchips and Meta's MTIA accelerator hardware.
A 10,000x increase in retrieval model capacity, replacing rule-based heuristics and isolated model stages.
Meta reports +6% recall improvement in the retrieval system and +8% ad quality improvement on selected segments.
Built to absorb the creative explosion from Advantage+ automation and generative AI tools.

What GEM adds

VERIFIED FACT. Meta's engineering post (November 10, 2025) describes GEM — the Generative Ads Recommendation Model — as its most advanced ads foundation model: LLM-scale, trained across thousands of GPUs. Documented results: a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in Q2 2025, and a Q3 architecture change Meta says doubled the performance benefit gained per unit of data and compute (Meta Engineering: GEM). GEM's knowledge transfers into the hundreds of smaller models that rank ads across Meta surfaces.

The enforcement layer is a different system — and it did move

VERIFIED FACT. Policy enforcement is documented outside the Andromeda/GEM posts, and those are the changes that affect whether an ad runs:

Ads are reviewed continuously and subject to re-review at any time (About Ads In Review, above).
Previously approved ads return to review after negative feedback or at random for accuracy checks.
Proactive enforcement operates at large scale: Meta's Integrity Reports state that in 2025 it removed over 159 million scam ads, 92% of them before anyone reported them (Meta Integrity Reports, H1 2026).
Policy standards keep updating on their own cadence — the Health and Wellness advertising standard shows change-log entries including December 27, 2024 and July 23, 2026 (Health and Wellness standard).

EVIDENCE-BACKED. Advertisers in regulated verticals report higher rejection rates in 2025–2026, and third-party trackers attribute it to better automated detection. Meta does not publish rejection-rate figures, so treat any specific percentage you see online as anecdotal. What you can verify yourself: the standards text, its change log, and your own account's ad-review outcomes in Business Support Home.

Pre-submission workflow

EVIDENCE-BACKED RECOMMENDATION — built on Meta's documented review behavior, not on Andromeda speculation

Keep a claim inventory: every claim you make anywhere, with the evidence file behind it. Variants inherit claims — control them at the source.
Write variants from the compliant claim set, not around it. Automation multiplies whatever wording you feed it.
Scan each variant's copy against the Meta profile before submission — the same 65 detectors, per variant.
Watch for re-review signals you control: hide/block/report rates on live ads pull ads back into review; creative fatigue raises them.
Re-check the standards' change log before each campaign flight.
Treat a clean approval as temporary. It was never a permanent license, and re-review is documented behavior.

What Optimus Pass checks (and what it does not)

OPTIMUS HEURISTIC — detector families, not Meta rules; scores are directional

Optimus Pass scans ad copy against 65 risk detectors across 10 platform profiles. For Meta campaigns, the relevant detector families are: prohibited and restricted claim language, overpromise framing, personal-attribute wording, restricted-vertical patterns, and implied transformation narratives.

It does not read Andromeda's or GEM's decisions, cannot predict delivery, and does not access ad accounts. Delivery performance and policy enforcement are Meta's systems; Optimus Pass only reviews the copy you paste in.

Scan every variant before you submit

Create a free verified account to run a scan: paste the ad copy, select Meta, and read the risk flags. No credit card is required. A clean scan does not guarantee approval — the decision is Meta's, at any time, on re-review too.

Run the free Ad Compliance scan →

Scaling creative volume?

The Agency plan covers 250 checks a month with shareable client reports — built for the reality that post-Andromeda accounts run dozens of variants, not one hero ad.

Compare plans →

Frequently asked questions

How does Meta ad review work?+

Meta states an ad may not be checked against all policies before it starts delivering, and that all ads are subject to re-review at any time. Review combines automated systems with human review: ads flagged by automation or belonging to restricted categories can be escalated, and live ads can be pulled back into review after negative feedback or at random for accuracy checks. The final approval decision is always Meta's.

What triggers a Meta ad rejection in regulated categories?+

Common policy-risk patterns include disease or treatment claims, before/after imagery, personal-attribute wording that implies the reader's health or financial status, absolute-outcome framing such as promises of instant or total results, and a mismatch between the ad and its landing page. These are the same families Optimus Pass scans for; the actual enforcement decision is Meta's, and no tool can guarantee it.

Did Andromeda change Meta's ad policies?+

No. Andromeda is Meta's ad retrieval system and GEM is its ads ranking foundation model — Meta's engineering posts describe them as delivery-performance systems, not enforcement systems. The rules that reject ads are the Advertising Standards, documented separately. What Andromeda changes is the environment your ads compete in, not what is allowed.

Why do advertisers say rejections got stricter in 2025-2026?+

Advertiser reports — especially in regulated verticals — describe more rejections and faster enforcement. That is observation, not a published Meta statistic: Meta does not publish rejection-rate figures, and it publishes no 'Andromeda policy'. The documented behavior that plausibly drives the experience is continuous review and re-review of live ads, plus proactive enforcement at scale — Meta reported removing over 159 million scam ads in 2025, 92 percent before anyone reported them.

Does creative volume affect compliance now?+

Indirectly, yes. Automation and generative tools let advertisers run far more variants, and Andromeda's retrieval models learn from people and ad content together. Every variant still has to pass the same Advertising Standards, so a larger creative library multiplies your compliance surface even though the rules did not change. The practical shift is from checking one hero ad to checking the whole set.

Can Optimus Pass keep my ads approved?+

No tool can promise that — Meta owns the decision. Optimus Pass scans ad copy against 65 risk detectors across 10 platform profiles, including the claim-language and personal-attribute patterns that drive most regulated-vertical rejections, and suggests lower-risk wording. Use it per variant before submission. It does not access your account, contact Meta, or influence delivery or enforcement outcomes.

References

Claim classes used on this page: VERIFIED FACT (documented directly by Meta), EVIDENCE-BACKED (grounded in those sources or in clearly-labeled advertiser observation), OPTIMUS HEURISTIC (Optimus Pass scoring conventions).

How Facebook Ad Review Works — And What Andromeda Changed